alterlab-pathml
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads the PathML framework and its associated library ecosystem from official package repositories. PathML is a recognized tool for medical image analysis.
- [COMMAND_EXECUTION]: The agent is instructed to use shell-based tools (Bash via uv/python) for data pipeline execution, dataset management, and deep learning model training. This includes interaction with HPC job schedulers like SLURM for batch processing of large-scale pathology datasets.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources, representing a theoretical surface for indirect injection attacks targeting the LLM during data processing.
- Ingestion points: Whole-slide images (e.g., SVS, NDPI), clinical metadata (CSV), and HDF5-formatted pathology datasets.
- Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded natural language content during metadata extraction or clinical record processing.
- Capability inventory: The skill has broad execution capabilities (Bash) and file system access (Write, Edit), which could be leveraged if an injection were successful.
- Sanitization: Standard data parsing is performed using NumPy, Pandas, and PathML, but no content-aware sanitization of metadata strings is implemented.
Audit Metadata